A local idea space: the value of personal and thematic proximity in academic research. Lukas Kuld. TEP Working Paper No

Transcription

1 A local idea space: the value of personal and thematic proximity in academic research Lukas Kuld TEP Working Paper No February 2017 Trinity Economics Papers Department of Economics Trinity College Dublin

2 A local idea space: the value of personal and thematic proximity in academic research Lukas Kuld February 20, 2017 Abstract While recent research has found no strong overall effects between colleagues in university departments, this paper shows a clear link between the success of individual research articles and local colleagues when their research is directly related. Using data from the CVs of around 1,000 academic economists, I study research links between department colleagues and their impact on citations. The novel focus on articles also addresses endogeneity concerns using differences in the scope of the effect for article quality and dissemination by publication type and peer group. The estimates show that articles in top 25 journals that draw on research of local colleagues receive significantly more citations than comparable work by the same authors. Conversely, the coauthor network is primarily correlated with low-profile journals and arguably reflect widely the authors field-specific standing. JEL-codes: A14 D83 I23 J24 O31 O32 O33 Keywords: Local research cluster, co-author network, tacit knowledge, scientific productivity, academic economic research Department of Economics, Trinity College Dublin, Dublin 2, Ireland; kuldl@tcd.ie I thank my PhD supervisor, John O Hagan, for extensive comments on several drafts of this paper and for funding, with the Department of Economics, of Interns Kate Hayes, Greg Mangan, Andrew Morrow, Sarah Mortell, Conor Parle and Michael Stone. I also thank Tara Bedi, Christiane Hellmanzik, Martina Kirchberger, Clemens Struck and Alan Walsh, and the participants of the Trinity College Dublin PhD working group for help and comments. The research is also generously supported by a Government of Ireland Postgraduate Scholarship from the Irish Research Council. 1

3 1. Introduction Access to research knowledge and new ideas is fundamental for the production of scientific knowledge. Within a scientific community, parts of this knowledge are created or shared in personal networks. The empirical literature is inconclusive though on the benefits from personal connections and their boundaries, in particular, the existence of localized peer effects. This paper expands and extends on recent empirical work on the decomposition of peer effects in science. In previous literature, the total annual article output of researchers is related to a change in the respective peer group, for example, a change in the department, co-author network or sub-field (see for example Agrawal et al. [2014], Borjas and Doran [2015], Azoulay et al. [2010], Waldinger [2010], Waldinger [2012] and Waldinger [2016]) 1. Most of these studies find no overall effect of researchers on the productivity of local colleagues or vice versa. With a focus on economists, two recent articles also observe no general localized peer effects within university departments over the last twenty years (Kim et al. [2009] and Bolli and Schläpfer [2015]) 2 which confirms the empirical evidence in other fields and time periods. The aggregation of different types of overall researcher profiles and different relations between researchers can, though, mask peer effects in subgroups, as university departments are typically not organized as singular research groups and, therefore, only parts 1 Agrawal et al. [2014] decompose the effects of a department hiring a star researcher in evolutionary biology into effects on overall research output by related and unrelated incumbents as well as new hires. While no positive effect is shown on incumbents overall, incumbents who work on similar research questions as the hired star increase their annual article output on average. The paper s main focus then is on the effect on new hires. Borjas and Doran [2015] study the potential negative effects resulting from the exodus of mathematicians after the end of the Soviet Union on the overall article output of collaborators left behind and previously geographically or thematically close researchers. Out of these groups, only former collaborators of highly productive emigrants appear to have been negatively affected. Azoulay et al. [2010] show the negative effects of the unexpected death of a scientific star on the productivity of collaborators in life science. Waldinger [2010] shows the importance of the quality of the supervisor on the lifelong productivity of PhD students in a sample of pre-world War II mathematicians. Waldinger [2012], in a later paper, finds no aggregated localized peer effects within German physics, chemistry or mathematics departments before the Second World War. Conversely, Waldinger [2016] finds negative short and long term impacts on the overall output of departments of the dismissed scientists. 2 Kim et al. [2009] study the effect of being affiliated with a top American economics or finance department. In their empirical setting, regressions of the individual overall annual publication record on the characteristics of the researcher include university fixed effects. These affiliation fixed effects are positive for the 1970s but are insignificant in the 1990s. As a consequence, they conclude that localized peer effects disappeared with the decrease of communication costs. Similarly, Bolli and Schläpfer [2015] conclude that German economists overall productivity did not profit on average from moves to new institutions between 2004 and

4 of a department s research will overlap. In this paper, I disaggregate from the overall departmental level to study peer effects in a more narrow definition of the research area. As a result, this paper examines the value of research proximity between individual colleagues. The contribution of this paper to the literature then is based on a novel focus on individual articles of researchers rather than the annual productivity of these researchers. 3 This perspective enables a more direct observation of the influence of peers, for example, if the research of a colleague is cited. In addition it allows a more detailed analysis of peer effects in interaction with other characteristics of the publication. 4 As a result and in contrast to earlier studies, this paper checks for local peer effects on individual article influence within specific research areas in the field of economics. A rich and new biographical data set on around 1,000 highly-cited economists makes this empirical study possible. The remainder of the paper is organized as follows. Section 2 describes the theoretical framework for peer effects in knowledge production. Sections 3 and 4 present the empirical framework to identify the influence of peers on research articles and the data used. Section 5 presents the empirical findings and Section 5 provides a discussion of the results and concludes. 2. Theoretical framework The literature on agglomeration commonly emphasizes the gains from clusters for individual productivity. Duranton and Puga [2004] categorize these into benefits from sharing (e.g. infrastructure), matching and learning. The sharing of infrastructure is less important in the empirical setting of economic research even if colleagues in the same area might be beneficial for cost-intensive projects. Following the increases in research collaboration and specialization (see e.g. Jones [2009]), matching of skills is of rising importance but only indirectly studied in this paper. The focus of this paper then is on the potential benefit for research projects from personal connections of the authors to, and hence learning from, other researchers, in particular, if the other researchers are not directly involved in the research project as authors but work on similar topics. 3 A similar classification for thematic proximity but on the level of individual researchers has been used for the analysis of peer effects of an incoming star researcher on incumbent scientists in Agrawal et al. [2014]. 4 However, this focus does not allow conclusions about the overall research output per author. Instead, the success of research articles with thematic links to colleagues or co-authors is compared to other work by the same author(s). 3

5 In the research and dissemination process of an economic research article, colleagues and former co-authors of the article s authors are two not mutually exclusive types of peers that potentially improve the article s quality, its visibility or other success factors. Subsequently, the network of non-local co-authors, local colleagues who are also co-authors and other local colleagues constitute the three main peer groups to be examined in this paper. Studying local colleagues, co-authorship within a university indicates particularly close working relations. However, the connection of colleagues who have not co-authored previously may possibly also add valuable complementary knowledge. In addition, the network of former co-authors is highly related to a researcher s characteristics such as status or specialization which also makes an empirical assessment challenging (see e.g. Ductor et al. [2014] and Kuld and O Hagan [2016]). The empirical estimation is, therefore, focused on the impact of local colleagues. To study benefits from peers formally, the influence that a research paper exerts on future research is used as the key definition of success in this paper. It is assumed that a research article s influence depends on the quality of the research, the dissemination process and the current and future interest in the research topic. Quality of a research article is seen as its potential contribution to the specific area regardless of the overall interest in the topic or dissemination and visibility effects. Dissemination denotes other factors that increase the influence of an article, in particular the reputation of authors and journals in the research area. The reputation gives a quality signal and increases visibility. In an efficiently working scientific field, quality and dissemination should be highly correlated but it is possible, for example, for a high-quality article to have little influence because of very low visibility. I i = f(q i, Dis i, Int i ) (1) Equation 1 then sets the influence of a research article i as a function of its quality, Q i, the dissemination process, Dis i, and the overall interest in the research area, Int i. In turn, quality and dissemination are potentially influenced by peers who are personally connected to the authors such as department colleagues and co-authors. A narrow definition of peer groups, with a resulting small number of peers, excludes a large influence of peers on the overall interest in the research area. As it is posited here that peers potentially can affect the quality of an article during its production, the quality of article i in Equation 2 is set as a function of the researchers in the neighborhood N i of the article, e.g. the departments of the authors. In addition, A i denotes the authors 4

6 and U i are characteristics of the authors institutes. Q i = h(u i, A i, g(a i, N i )) (2) The influence of peers depends potentially on the form of the relationship between the authors and the peers which is expressed by g(a, N). For example, this paper focuses on research proximity and co-location and argues that peers are more helpful if the research is closely related and if the researchers co-locate or co-author during the research process. Two main arguments suggest potential quality effects from close personal contact to thematically related researchers. First, most knowledge to produce research is tacit and not easily accessible but apprehended, i.e. transmitted and created, by personal contact (see Polanyi [1958]). 5 Second, new ideas and knowledge are at first circulated in personal connections or created within personal connections. 6 As a result, personal connections lead to a more efficient research process which is particularly important in a dynamic research field: In the trade-off between maturing a research project and ensuring priority by publication (Bobtcheff et al. [2016]), the help and critique of colleagues lead to a more efficient maturation process and, therefore, a higher quality at the time of publication. The importance of tacit knowledge may change over the career of a researcher. Established researchers possess most of the relevant tacit knowledge within their discipline. However, this shared tacit knowledge facilitates, subsequently, the creation and discussion of new ideas. Therefore, interactions with other researchers are more fruitful if the research focus is similar. As a consequence, the importance of personal connections is expected to be at its highest at the very beginning of a career. Lower but stable effects are expected in later career stages. 7 5 While all knowledge is somewhat tacit or rooted in tacit knowledge, tacit knowledge is defined here as knowledge that is not easily codified and made accessible to other scientists, for example, the appreciation of a mathematical theory by scientists before experimental observation and the theory s relation to the observation thereafter (Polanyi [1958], p.60). 6 See e.g. Borowiecki [2013], Hellmanzik [2010], and Mitchell [2016] for evidence for this concept in artistic production. 7 As argued by Kim et al. [2009], cheap air travel and the Internet can change the setting of this knowledge transfers and creation as it allows for an easier connection with distant colleagues. These connections though have a higher cost and are less likely. In addition, lower communication costs, higher mobility of scientists and the increased specialization of research, can threaten departments as a common social and work space. As a result, the departments role in enabling new research connections and mutual assistance between researchers is potentially weakened. Through the Internet and air travel, scientists can find a better matching of skills with researchers at distant universities and subsequently stay clear of department colleagues. This has potentially negative effects for less experienced researchers without a network and recalls the image of Schopenhauer s republic of geniuses (as opposed to the republic of scholars ) in which according to Nietzsche One giant calls to the other across the bleak intervals of time and the conversation of the great minds goes on 5

7 Besides quality effects, personal connections can impact also on the dissemination process more directly. First, colleagues and former co-authors are potentially more aware of and influenced by ongoing research by the authors. Second, connected researchers can influence the wider dissemination process, for example, as journal editors or through conferences. Furthermore, a high number of personally connected researchers, in particular former co-authors, working in a given sub-field may indicate the authors prominent standing in the sub-field. From the perspective of a potentially influenced researcher, reputation is a quality signal which reduces the effort for quality screening. In addition, aligning with influential research, which is signaled by reputation, makes research more likely to be of interest to (or commensurable with) future research. Therefore, the reputations of the authors and of their affiliations and peers, and the type of journal impact positively on the dissemination of a research article. In turn, the journal depends among others on the quality of the article. Taken together, dissemination can be set as Dis = Dis(A, U, N, J(Q)). As a consequence, peers in the neighborhood N can impact an article s influence through quality and dissemination which leads to the following overall relation between influence and peers in Equation 3, ceteris paribus. ( I(N) = f Q (., N ), Dis ( Q(., N),., N ) ), Int (3) The empirical framework below outlines, first, an estimate of the overall impact of peers on the success of an article and, in a second step, a differentiation of peer effects into dissemination and quality effects. Although less important for individual influence, a distinction between quality and dissemination effects is helpful to assess the social benefit of peer effects and related policies. For the advancement of a scientific field, quality effects are arguably more important than individual dissemination. In addition, dissemination advantages are arguably less intrinsically dependent on co-location. undisturbed by the mischievous, noisy dwarfs who creep among them. (Nietzsche [1954, 1E 1872], p.269)). Analyzing research articles that are explicitly influenced by a colleague allows us to assess peer effects without being as strongly affected in the analysis by this development. 6

8 3. Empirical framework Empirical model The influence of any article is measured in this paper by the number of citations (see e.g. Bornmann and Daniel [2008] or Osterloh and Frey [2014] for a discussion on impact and citation counts) 8. The empirical identification of peer effects then is based on their impact on the total sum of citations received per article (I). Next, research links to local colleagues are estimated by classifying articles on whether they refer to articles by recent colleagues of the authors. Possible endogeneity issues of proximity to eminent researchers are addressed, first, by focusing on linked research instead of university affiliation and using individual and university dummies, second, by comparing publication sub-groups with different scopes for quality maturation and dissemination effects, and, third, by contrasting estimates for the intrinsically endogenous network of former co-authors. More detail is given below. As a first step, the overall impact of peers referenced, colleagues and co-authors, in an article on the influence of the article is estimated by: log ( E(I U, A, T, X, N) ) = U + A + T + Xγ + Nβ (4) Equation 4 shows the estimation of the mean of citations received, I, as an exponential function of university U, author A and time fixed effects T, a vector of control variables X and indicator variables for a reference in the article to a colleague or former co-author: N = (CoAuthor, Colleague). The key variables of interest are the peer variables in N and these will be discussed first. Peer variables The variables in N = (CoAuthor, Colleague) denote articles that reference a recent coauthor and/or colleague of the authors. Figure 1 sketches the applied measurement of personal and thematic links. For illustrative purposes, we assume that three new articles are written by authors at two different affiliations: Square University and Triangle University. Each article refers to five existing articles which defines their research area 8 Bornmann and Daniel [2008] give an overview of studies on the relation of citation counts and impact. While the citation behavior varies between researchers, the authors conclude that citation counts are generally a valid measure of impact. 7

9 Figure 1: Thematic and personal relations in an article citation graph. Old1 Old2 Old3 Old4 Old5 Old6 Old7 Old8 Old9 Old10 Authors at Square Universiy Triangle University Other affiliation New Article 1 New Article 2 New Article 3 relatively. For example, New Article 1 and New Article 2 have a similar research focus based on their references while New Article 3 has a different topic. Of the three new articles, the only article for which the thematic and personal links overlap is Article 1, i.e. the article cites articles that are written by recent colleagues of the authors. In consequence, New Article 1 is the only article of the three that is classified as treated with the respect to the variable Colleague. The further empirical analysis is centered on the effect of this treatment. Formally, S i is constructed as the set of authors referenced in article i after excluding article references that are (co-)authored by an author of the article. 9 Next, this set of individuals is compared with the list of eminent economists in our sample 10 who have a known personal connection to an author around the time of production, i.e. are in the neighborhood N x i of the article. N CoAuthor i and N Colleague i are defined as the sets of economists in the sample who have shared an affiliation or co-authored an article one to five years prior to the publication with at least one of the authors of the article i. The 9 This is varied in robustness check. The exclusion follows the argument that the co-authors should have complementary knowledge on the topic. 10 The sample of 967 eminent economists as described in the data section. 8

10 time frame is chosen to reflect the likely production period of an article. Subsequently, two variables are constructed that indicate the articles in which at least one colleague { or former co-author } from this time span is referenced: Colleague = 1 A, with A = i N Colleague i S i and CoAuthor = 1 B, with B = { i Ni CoAuthor S i }. Control variables The vector X contains a series of further characteristics of the article and its authors. First, the mean age of the authors and the squared mean age. This is included to account for career effects on productivity. Second, the number of affiliations, the number of authors and whether the authors are listed alphabetically is used to complement the individual and university dummies. A non-alphabetical ordering is unusual in economics 11 and can indicate a different background of the authors or authors added without a full contribution which would overestimate the number of authors. Third, the number of references and its squared value are included. More references can indicate a bigger project, more interest in the research area or increase the visibility of the publication independently. 12 The number of pages is not significant if the number of references is used and, subsequently, not used in the estimation. Additional specifications introduce a control variable for the total number of eminent economists cited. Since the sample of economists was selected based on citations in economic journals, a high number of cited eminent economists indicates a research field that attracts a high interest by economic researchers. The interest in the topic could in turn cause a positive effect of peers referenced. However, as shown below, the main estimates for the influence of peers are not changed significantly after the introduction of this control variable. This variable is not included in the main regression models as it is difficult to rule out connections between citing researchers. Finally, the number of selfreferences is counted to indicate prior experience in the area: NumSelfReferences i = M i, where M i is the set of references in i to articles by an author of i. Prior publications could indicate a higher visibility and be helpful in the research process. 11 In the time span studied, 92 per cent of articles published in AER, Ecta, JPE, OJE and REStud with more than one author list the authors alphabetically. 12 For example, on-line databases make it possible to search citing papers. Therefore, the more references an article lists the more such searches include the article. 9

11 Identification The empirical analysis focuses on the impact of research links to local colleagues. Endogenous peer effects, i.e. how the effects are driven by the influence or quality of linked research knowledge, and exogenous peer effects, i.e. the exogenous characteristics of these colleagues, are not explicitly distinguished (see Manski [1993]). Therefore, the main focus is on assessing whether correlated effects are expressed in the estimates. As outlined above, this paper addresses the endogeneity concerns through three channels. First, the focus on linked research articles allows the use of university and author fixed effects to address the endogenous selection into stronger research departments and differences in individual research performances. The fixed effects are varied and additional dummies for the specific group of authors and journal are introduced. Second, the estimates for local colleagues are compared to equivalent estimates for former co-authors. The network of former co-authors is by definition a strong indicator of past research performance. References to the research of former co-authors can, therefore, indicate, for example, previous field-specific research. If the two peer groups express such similar issues, then an equivalent estimation process of this peer group and local colleagues leads to similar estimates. Strongly different estimates support that the estimates are not dominated by common correlated aspects of these peer groups or the estimation process. In addition, regressions for sub-groups, such as authors with less prominent research records or at particularly prestigious universities show if the estimates were caused by specific outliers or correlations. Third, separate estimates for high- and lower-profile journals reflect the outlined theoretical channels for the impact of linked colleagues. As a result, estimates can agree with the theoretical expectations and at the same time contradict possible correlated effects that would, for example, upward bias all estimates. Seeing diverging estimates supports, thereby, the plausibility of the hypotheses posited. The link between the theoretical channels and the analysis by journal profile is based on two assumptions. First, research that is published in high-profile journals should be more challenging to the authors. Therefore, these publications should have more scope for quality benefits from the interaction with peers. Second, low-profile publications in journals that are commonly given less attention should benefit relatively more from increased attention due to the authors or their peers standing in the specific field. A journal with a high reputation gives a strong quality signal and increased visibility independently of the authors characteristics. Positing reputation as the main factor in the dissemination, this implies that articles in top journals should benefit less relatively 10

12 from the reputation of their authors, affiliations, or peers than publications in lowerranked journals. In addition, direct citations by personally connected researchers have a larger relative impact on the citations of less cited articles. 13 For the estimation by journal profile, Colleague and CoAuthor are interacted with the relative standing of the article s journal: The variable T J denotes a publication in a top 25 economics or finance journal. 14 This then gives Equation 5 as log ( E(I U, A, T, X, N, J) ) = U + A + T + J + Xγ + (N, N T J)β (5) As above, U, A and T are university, author and time fixed effects. J are journal fixed effects; X is a vector of control variables. The vector of peer variables takes now the form (N, N T J) = (CoAuthor, Colleague, CoAuthor T J, Colleague T J) and, in consequence, β is the 4-dimensional vector: β = (β 1, β 2, β 3, β 4 ) T. Three main extensions to this regression model are presented below. First, department colleagues are distinguished by whether or not they co-authored prior to the publication. This is used to repeat the estimations with three instead of two peer groups, i.e. the peer vector N includes then non-local co-authors, co-authors within the department and other department colleagues. Second, publications of younger researchers are identified. As outlined in the theoretical section, it is expected that peer effects vary over the career span of a researcher, being the strongest at the very beginning. To address this, the variable M axage5 denotes articles of which no author finished their postgraduate studies more than five years prior to publication. The large majority of the economists in the sample has non-permanent positions and shows different publication patterns with a higher share of single-authored papers during the first five career years (see Kuld and O Hagan [2016]). The peer variables are, subsequently, interacted with the age variable to estimate peer effects at the different age groups. Third and finally, the estimation is repeated for different sub-groups such as more and less eminent economists and lowand high-profile journal publications. This is intended as a robustness check and an indication on how the results hold for less productive researchers outside the sample 13 A wide interest in the specific research area should equally impact high- and low-profile publications. The importance of a publication s visibility for citations received is, for example, suggested by Feenberg et al. [2015] and Judge et al. [2007]. 14 To rank the 255 economics journals, the set of all research articles between 1996 and 2014 as described in the data section is used. First, all citations are divided by the yearly median. Then, the mean of these adjusted citations per article is used to rank the journals. Finally, the first structural break in the mean of citations at 25 was used to classify the publications. 11

13 Figure 2: The affiliations of the authors in the sample Longitude Latitude Articles in sample Under to to 200 Over 200 studied. 4. Data The empirical analysis is carried out using publications by the most cited academic economists between 1996 and The construction of this data set started with the economic journals listed by Kalaitzidakis et al. [2011]. These over two hundred journals were supplemented with a number of other, highly ranked journals in Ideas RePEc, e.g. the relatively recent AEA American Economic Journals, which brings the total number to 255 journals. Next, the authors of these articles were ranked by the number of citations received. From this exercise, a total of 967 economists were chosen based on work published in the period 1996 to 2013 and most highly cited in this period. All of these authors were ranked among the top 1,300 economists. The reasons that not all 1,300 top ranked economists are included is that CVs were not available or that their name details were not individual enough to be confidently attributed to a single economist. As an illustration of the spatial distribution of the universities studied, the affiliations of the selected economists are shown in Figure 2. While many countries are observed, there is still a strong concentration on North-American universities. Within the sample, economists from Harvard and Berkeley account for over ten per cent of the article output 12

14 and the ten universities with the highest share of articles are in the USA. Education is even more concentrated; twenty-one per cent of the sample economists hold a PhD from Harvard or MIT and the top six universities account for forty-four per cent of the doctorates. For the economists selected, complete information on the research career from the undergraduate studies onward was compiled using on-line CVs and the encyclopedia Who s Who in Economics (Blaug and Vane [2003]). This was complemented using name searches on Scopus to retrieve additional publication data. In most cases, the author s Scopus profile contains a complete list of the author s available publications. The data sets used in the regressions below are restricted to publications between 1996 and This yields 28,901 research articles that include at least one sample economist. The main regressions below use a more restricted data set of which all authors are in the sample. In addition, articles in the Journal of Economic Literature or articles that cite less than five or more than 80 other articles are excluded. This main data set, then, contains 7,291 research articles. Figure 3: Citations received per article Colleagues referenced 2000 Co authors referenced 0 0 Articles 1 2 Articles Citations Citations (a) Citations by number of colleagues refer-(benced. Citations by number of co-authors referenced. Notes: The background shows the citations received of 28,901 economic research articles between 1996 and 2014 of which at least one author are eminent economists as described in data section. The vertical lines show the means of citations received by number of eminent colleagues/co-authors referenced. The plots are truncated at 200 citations. Before the estimation of the regression models, Figure 3 shows the citation averages of articles by the number of referenced eminent colleagues and former co-authors. This indicates in a preliminary, descriptive way a positive relation between the citations received by an article and the authors personal connections to eminent researchers with 13

15 related prior publications, something that is tested more formally below. The graph on the left in Figure 3 shows that on average every additional eminent colleague referenced is associated with an increase of slightly over 20 citations received per article with up to four colleagues referenced. In contrast, co-authors are only associated with an increase in citations for the first two co-authors referenced (graph on the right in Figure 3). For both categories, higher counts of peers referenced than depicted are very uncommon. 5. Empirical results This section reports on the empirical estimation applied to the models looked at in Section 4, namely the estimates of the impact of local peers on the success of individual research articles by type of journal. 15 Formally testing, then, the impact of peers on individual articles, Table 1 shows that linked colleagues impact strongly and significantly on the influence of a high-profile article. This holds throughout the different specifications presented in the table which all correspond to Equation 5 in Section For example, in specification (3), the impact of colleagues who are referenced in a high-profile publication is estimated to increase the number of citations received on average by 35 per cent. 17 The estimates in Table 1 also show the distinct effects of the two peer groups: Colleagues impact heavily on the influence of articles that are published in high profile journals while co-authors are associated with a strong increase in citations to publications in journals with a lower profile. Following the preceding discussion, this indicates an impact of department colleagues on the quality of challenging research while linked co-authors may indicate advantages in the dissemination of less central publications. Reassuringly, the estimated parameters of the other co-variates do not surprise. The negative estimates for selfreferences and universities issue from the overlap with other references and the number 15 All main estimates are derived using quasi-poisson regressions using the natural logarithm as the link function. The literature on citation data is inconclusive on estimation methods but none of the models used changes the key results qualitatively. Quasi-Poisson regressions account for the count nature of citations, the observed over-dispersion and is more robust than negative binomial regressions given the high number of dummies. The estimates are robust to different estimation models such as a linear model using a log-transformed variable or negative binomial models. The estimation is carried out using the glm function in R. The packages multiwayvcov (multi-way clusterrobust variance estimation as suggested by Cameron et al. (2011)) and lmtest (hypothesis testing) are used for standard error correction. 16 Here, the impact of colleagues and co-authors is separately estimated for publications in top 25 and other journals. 17 The percentage effect for β = is calculated as (e ) 100 = 35.26, see (Wooldridge [2010], p. 726) 14

16 Table 1: Impact of peers by journal profile Dependent variable: SumReceivedCitations (link: log) (1) (2) (3) (4) (5) CoAuthor:Top25Journal (0.081) (0.097) (0.099) (0.163) (0.108) CoAuthor:OtherJournal (0.102) (0.095) (0.081) (0.136) (0.084) Colleague:Top25Journal (0.062) (0.079) (0.080) (0.119) (0.089) Colleague:OtherJournal (0.104) (0.092) (0.079) (0.108) (0.075) Top25Journal (0.054) (0.053) (0.052) (0.072) NumAuthors (0.080) (0.366) (0.384) (0.333) NumberUnis (0.072) (0.079) (0.206) (0.474) (0.242) MeanAcademicAge (0.007) (0.012) (0.012) (0.137) (0.013) MeanAcademicAge (0.000) (0.000) (0.000) (0.000) (0.000) NumReferences (0.002) (0.002) (0.002) (0.002) (0.002) NumSelfReferences (0.006) (0.007) (0.007) (0.009) (0.007) OrderAuthors (0.054) (0.068) (0.070) (0.181) (0.063) Year dummies Yes Yes Yes Yes Yes Individual dummies No Yes Yes No Yes Authorgroup dummies No No No Yes No University dummies No No Yes Yes Yes Journal dummies No No No No Yes N (df) 6958 (6928) 6958 (5993) 6958 (5730) 6958 (4904) 6958 (4884) Pseudo-R Notes: This table reports estimated coefficients from quasi-poisson regressions with standard errors in parentheses. Standard errors are clustered at the journal-year level except for model (4) which is clustered at the author group level (variables as described in Table A.1). Each observation is an article of which all authors are in the sample described of 967 eminent economists. : Universities with more than two observations. p<0.1; p<0.05; p<

17 of authors and the real estimated effect is, therefore, in both cases the sum of the parameters. For comparison, Table A.2 (appendix) shows the estimates of the overall impact of peers on individual articles without the differentiation by journal profile: Local colleagues are associated with a significant increase in citations received for all specifications. The overall relation of the success of individual articles and linked former co-authors is less clear as the estimated effects are smaller and the variance higher. So far, colleagues who are also co-authors were included in both categories. When estimating the effect of this group on its own (Table 2), it shows that their impact is similar to non-local co-authors once individual dummies are introduced. Importantly, the introduction of individual dummies leads to opposite changes in the estimated impact: a decrease for co-author-colleagues and an increase for other colleagues. The decrease due to individual dummies suggests that co-author-colleagues indicate very productive individuals. However, this peer group is less correlated with higher citation counts once the averages of the authors are considered. Conversely, the use of individual dummies shows that local colleagues outside of the authors co-author network impact positively on high-profile research. Research projects are significantly more influential than other research by the same authors if they relate to research by this complementary group of colleagues. Robustness checks and extensions Tables 1, A.2 and 2 show estimates that are based on publications of which all authors are in the sample of eminent economists. This raises questions on the representativeness for all publications of these authors and for publications of other less prominent authors. For example, the limitation to publications of which all authors are sample economists leads to an under-representation of multi-authored papers. To address representativeness within the sample of top economists, Table A.3 (appendix) repeats the regressions with all publications of the sample economists. While this introduces new problems due to unknown authors, the estimates are still strong and significant for the impact of colleagues on high-profile and the impact of co-authors on low-profile publications. This shows that the results are not limited to the more restrictive sample or caused by the over-representation of single-authored articles. In addition, Table A.6 shows the parameters for colleagues and co-authors estimated separately which rules out that the estimated effect issues from the correlation between the two variables. Table A.7 shows the estimates of the regressions presented in Table 1 while also controlling for the overall 16

18 Table 2: Impact of co-authors, colleague-co-authors and other local colleagues Dependent variable: SumReceivedCitations (link: log) (1) (2) (3) (4) (5) CoAuthor:OtherUni:Top25Journal (0.148) (0.155) (0.169) (0.142) (0.155) CoAuthor:OtherUni:OtherJournal (0.118) (0.175) (0.126) (0.178) (0.118) Colleague:NoCoAuthor:Top25Journal (0.067) (0.076) (0.074) (0.125) (0.079) Colleague:NoCoAuthor:OtherJournal (0.110) (0.084) (0.076) (0.113) (0.078) Colleague:CoAuthor:Top25Journal (0.133) (0.146) (0.153) (0.206) (0.162) Colleague:CoAuthor:OtherJournal (0.160) (0.186) (0.154) (0.181) (0.127) Top25Journal (0.053) (0.051) (0.050) (0.070) NumAuthors (0.091) (0.440) (0.474) (0.423) NumberUnis (0.070) (0.080) (0.177) (0.407) (0.206) MeanAcademicAge (0.007) (0.012) (0.012) (0.150) (0.012) MeanAcademicAge (0.000) (0.000) (0.000) (0.000) (0.000) NumReferences (0.002) (0.002) (0.002) (0.002) (0.002) NumSelfReferences (0.006) (0.007) (0.007) (0.009) (0.007) OrderAuthors (0.079) (0.088) (0.095) (0.193) (0.087) Year dummies Yes Yes Yes Yes Yes Individual dummies No Yes Yes No Yes Authorgroup dummies No No No Yes No University dummies No No Yes Yes Yes Journal dummies No No No No Yes N (df) 6958 (6926) 6958 (5991) 6958 (5746) 6958 (4917) 6958 (4900) Pseudo-R Notes: This table reports estimated coefficients from quasi-poisson regressions with standard errors in parentheses. Standard errors are clustered at the journal-year level except for model (4) which is clustered at the author group level (variables as described in Table A.1). Each observation is an article of which all authors are in the sample described of 967 eminent economists. : Universities with more than two observations. p<0.1; p<0.05; p<

19 number of eminent economists cited. The estimates show that the estimated peer effects are not driven by the interest in the topic as evidenced by eminent colleagues working on it. However, the main specification does not include this control variable since a personal connection between the researchers cannot be ruled out. To indicate the effects on less eminent researchers, Table A.4 (appendix) repeats the estimation of specification (3) in the main Table 1 leaving out publications by the most cited 100 (column 1) to 500 (column 5) economists. The literature stresses the importance of star researchers (see e.g. Agrawal et al. [2014] and Oettl [2012]) on the productivity of peers, hierarchical effects on less productive researchers would, therefore, lead to higher estimates. Conversely, these slightly less cited researchers have on average less eminent sample economists in their departmental or co-author network which would lead to lower effects if the peer effects of stars are larger. Overall, the estimates in Table A.4 do not show a clear trend, while the observed pattern persists that local colleagues strongly impact on high-profile publications. Next, a distinction of the peer effects by age is presented in Table 3. Confirming the expectation outlined in the theoretical framework, the effects of co-authors or colleagues are estimated to be stronger during the first five years following the award of a PhD. Naturally, the number of co-authors is lower for younger researchers which is reflected in higher standard errors. Importantly, colleagues and co-authors persist to impact on articles in the same pattern as in Table 1 even after the initial career stage. Conclusion of empirical results As discussed previously, this paper focuses on individual articles instead of the overall article output of a researcher. Overall, the co-author network might impact more positively on individual productivity, in particular as direct co-authorship is not studied here (see e.g. Azoulay et al. [2010], Borjas and Doran [2015], Ductor et al. [2014] and Ductor [2015]). In particular, learning through co-authorship may lead to an overall higher research performance which does not show in the comparison of individual research projects. In addition, the co-author network is arguably a stronger reflection of a researcher s past productivity than local colleagues, if the affiliation is controlled for. In turn, this lower dependence on past performance reaffirms any estimated positive impact of local colleagues outside the researcher s co-author network. 18 This causality in the estimated impact of local colleagues is further supported by the estimated pattern 18 The negligible impact of university dummies on the estimated effect of colleagues agrees with the posited low dependance. 18

20 Table 3: Impact of peers by age and journal profile Dependent variable: SumReceivedCitations (link: log) (1) (2) (3) (4) (5) CoAuthor:Top25Journal:MaxAge (0.215) (0.245) (0.237) (0.424) (0.243) CoAuthor:Top25Journal:OtherAgeProfile (0.082) (0.105) (0.108) (0.167) (0.122) CoAuthor:OtherJournal:MaxAge (0.219) (0.209) (0.238) (0.305) (0.260) CoAuthor:OtherJournal:OtherAgeProfile (0.112) (0.103) (0.087) (0.144) (0.089) Colleague:Top25Journal:MaxAge (0.123) (0.126) (0.133) (0.200) (0.146) Colleague:Top25Journal:OtherAgeProfile (0.069) (0.085) (0.084) (0.132) (0.091) Colleague:OtherJournal:MaxAge (0.139) (0.126) (0.135) (0.182) (0.145) Colleague:OtherJournal:OtherAgeProfile (0.127) (0.109) (0.091) (0.125) (0.085) Top25Journal (0.054) (0.054) (0.053) (0.072) MaxAge (0.086) (0.087) (0.087) (0.142) (0.092) NumAuthors (0.080) (0.370) (0.388) (0.336) NumberUnis (0.072) (0.078) (0.208) (0.481) (0.246) MeanAcademicAge (0.009) (0.015) (0.014) (0.140) (0.016) MeanAcademicAge (0.000) (0.000) (0.000) (0.000) (0.000) NumReferences (0.002) (0.002) (0.002) (0.002) (0.002) NumSelfReferences (0.006) (0.007) (0.007) (0.009) (0.007) OrderAuthors (0.054) (0.068) (0.070) (0.181) (0.063) Year dummies Yes Yes Yes Yes Yes Individual dummies No Yes Yes No Yes Authorgroup dummies No No No Yes No University dummies No No Yes Yes Yes Journal dummies No No No No Yes N (df) 6958 (6923) 6958 (5988) 6958 (5725) 6958 (4899) 6958 (4879) Pseudo-R Notes: This table reports estimated coefficients from quasi-poisson regressions with standard errors in parentheses. Standard errors are clustered at the journal-year level except for model (4) which is clustered at the author group level (variables as described in Table A.1). Each observation is an article of which all authors are in the sample described of 967 eminent economists. : Universities with more than two observations. p<0.1; p<0.05; p<

21 in which the effect is limited to high-profile research. First, a strong quality effect on low-profile research is unlikely given the sample of eminent researchers. Second, confounding factors that influence both publication types are ruled out. Third, the contrast to the co-author estimates rules out confounding factors that are similarly correlated with both groups. On the other hand, besides dissemination effects, additional arguments suggest a relatively higher importance of co-authors for lower-profile publications: First, in comparison to the individual average, co-authors are an expression of the authors research record in different sub-fields and the indicated ability and reputation are, therefore, field-specific. As a consequence, co-authors have a higher expected impact on field journals which are typically lower ranked. Second, co-authors may be important for the generation of new ideas but that the quality of high-profile publications depends more on the efficiency of the maturation (Bobtcheff et al. [2016]) process which may be more strongly influenced by local colleagues Conclusion This paper argues that personal connections to thematically close researchers are crucial for the production and dissemination of research. To test this proposition, economic research articles are classified by reference to local colleagues or former co-authors. Furthermore, the publications are distinguished by the standing of the journal to separate a researcher s more and less high-profile research. Finally, the number of citations received is taken as an estimate for an article s influence or success. In this framework, thematic proximity to research of local colleagues and former coauthors is shown to increase the average number of citations received significantly. The influence of co-authors is primarily to low-profile publications which might be due to dissemination effects of the co-author network. However, the findings do not allow to confidently distinguish between former co-authors as a reflection of the authors research record and a direct impact of co-authors on the article. The estimates show, though, that high-profile research benefits strongly when it can also draw on the research of local colleagues outside the co-author network. 19 Alternatively, different researchers rely on different peer groups and publish in different journals. However, strong effects are estimated for colleagues throughout different subgroups of researchers in Table A.4 which makes this explanation unlikely. Finally, reverse causality that colleagues are systematically cited in expected stronger publications can be excluded as a strong factor in the estimates. In Table A.5, local colleagues are not cited significantly more often in top journal publications when individual dummies are used. 20